For better or worse, the value of code itself seems to be dropping precipitously, to be replaced by measures like how well an LLM can understand the codebase (
An experiment in vibe coding — @nolanCLAUDE.md,AGENTS.md) or how easily it can test its “fixes” (unit/integration tests).
This was a timely post. I also played with vibe coding during the holidays, also to build an app for my wife. And I came to just about the same conclusions 🙂
I’ve used AI at work ever since GitHub Copilot became available, and played with different options as they became available. Yet I had never built an app from scratch using only AI, without looking at the code. It was an interesting experiment.
Claude Opus 4.5 got me to a working app very quickly. I had to take forced breaks when running into AI usage limits, which isn’t something I’m used to, but that’s probably to be expected when fully relying on AI to do the work.
I’m used to the dopamine hit I get when finding the solution to a bug after debugging for a while. This was a different feeling ; even though the AI was doing all the coding work, I felt creative anyway, having to explain and clarify my ideas, and iterate based on the feedback from the AI.
After a couple of hours, the app felt like 90% of the way there, it was very impressive.
At that point I decided to hit pause on the creation work, and spend some time testing the app. That’s when the problems started. I ran into bugs, performance issues, security issues, and missing functionality.
I asked AI to fix some of those bugs, but quickly realized this wasn’t efficient. Instead, I asked AI to set up unit tests, code formatting rules, linting, e2e tests, GitHub action workflows, and maybe most importantly, Claude and Copilot rules so agents would respect these requirements.
Once that was set up, I used Claude Code agents and GitHub Copilot to work on Pull Requests to fix all the problems in the code. This proved to be a better approach.
All in all, it was a positive, frustrating at times, humbling experience. While it was a good demonstration of the quality you can get when vibe coding, it was also a good insight into the potential future of small hobby apps.
It was also a good reminder that I like coding! I missed being able to work out on those bugs on my own. As @davidcelis said best:
Writing code is fun and helps me improve my skills. Solving problems keeps my mind sharp and helps me learn even more.
Writing Code Is Fun
2 responses
@jeremy Great post that matches my experience! Using Claude Code without tests is a great way to build a complete mess. And it _can_ fix accessibility, performance, and security issues, but you have to ask it (which is annoying).
I actually used Claude Code to find a low-severity security bug in the vibe-coded app I built (bug of course due to Claude's prior work). I just said "do a security audit" and it found it. These tools _can_ get work done, but they need guidance and guardrails.
Thanks for sharing the experience! I’m glad it reminded you that you love to code! Even though it sounds like the AI coding was very impressive, your post reminds me in a good way, that the human element is still needed in the creation process. Keep up the good work, and I hope you continue to have fun with coding.